On the Relevance of Sparsity for Image Classification

In this paper we empirically analyze the importance of sparsifying representations for classification purposes. We focus on those obtained by convolving images with linear filters, which can be either hand designed or learned, and perform extensive experiments on two important Computer Vision problems, image categorization and pixel classification. To this end, we adopt a simple modular architecture that encompasses many recently proposed models. The key outcome of our investigations is that enforcing sparsity constraints on features extracted in a convolutional architecture does not improve classification performance, whereas it does so when redundancy is artificially introduced. This is very relevant for practical purposes, since it implies that the expensive run-time optimization required to sparsify the representation is not always justified, and therefore that computational costs can be drastically reduced.

Fethallah Benmansour, Matthew Brown, Pascal Fua, Germán González Serrano, Vincent Lepetit, Roberto Rigamonti, Engin Türetken

In this paper we analyze empirically the importance of sparsifying representations for classification purposes. We focus on those obtained by convolving images with linear filters, which can be either hand designed or learned, and perform extensive experiments on two important Computer Vision problems, image categorization and pixel classification. To this end, we adopt a simple modular architecture that encompasses many recently proposed models. The key outcome of our investigations is that enforcing sparsity constraints on features extracted in a convolutional architecture does not improve classification performance, whereas it does so when redundancy is artificially introduced. This is very relevant for practical purposes, since it implies that the expensive run-time optimization required to sparsify the representation is not always justified, and therefore that computational costs can be drastically reduced.

2012

Object Classification and Detection in High Dimensional Feature Space

Charles Dubout

Object classification and detection aim at recognizing and localizing objects in real-world images. They are fundamental computer vision problems and a prerequisite for full scene understanding. Their difficulty lies in the large number of possible object positions and the appearance variations of object classes. This thesis improves upon several classical machine learning algorithms, enabling large computational gains in high dimensional feature space. A common trend in machine learning and computer vision research is to go large scale. In particular, the advent of huge datasets mined from the Internet, and the combination of multiple feature sources have considerably broadened the applications of computer vision. Tasks which were thought impossible a few years ago, such as human action recognition or pose estimation, automatic outdoor navigation, etc., now seem within reach. This dissertation is divided into two parts. The first one deals with the efficient training of a classifier or detector based on a large number of feature extractors, outside the control of the learning algorithm, and therefore of unknown suitability to the task at hand. More precisely, this part presents two kinds of strategies to accelerate the training of Boosting algorithms in such a context: (a) a method to better deal with the increasingly common case where features come from multiple sources (e.g. color, shape, texture, etc., in the case of images) and therefore can be partitioned into meaningful subsets; (b) new algorithms which balance at every Boosting iteration the number of weak learners and the number of training examples to look at in order to maximize the expected loss reduction. Experiments in image classification and object recognition on four standard computer vision datasets show that the adaptive techniques we propose outperform both basic sampling and state-of-the-art bandit methods. The second part deals with linear object detectors, currently the most popular class of detection systems, encompassing template matching, deformable part models, poselets, convolutional neural networks (which internally use linear filters), etc. The main bottleneck of many of those systems is the computational cost of the convolutions between the multiple rescalings of the image to process and the linear filters. We make use of properties of the Fourier transform and clever implementation strategies to obtain a speedup factor proportional to the filter size, both while training and at test time. We also introduce a few modifications to the original Deformable Part Model (DPM) of Felzenszwalb et al. improving its detection accuracy. The gains in performance are demonstrated on the well-known Pascal VOC benchmark, where an increase by one order of magnitude in the speed of said convolutions, and an average improvement of 15% in the accuracy of the detector are established.

Programme doctoral en Informatique, Communications et Information2013

Multiscale Centerline Extraction Based on Regression and Projection onto the Set of Elongated Structures

Amos Sironi

Automatically extracting linear structures from images is a fundamental low-level vision problem with numerous applications in different domains. Centerline detection and radial estimation are the first crucial steps in most Computer Vision pipelines aiming to reconstruct linear structures. Existing techniques rely either on hand-crafted filters, designed to respond to ideal profiles of the linear structure, or on classification-based approaches, which automatically learn to detect centerline points from data. Hand-crafted methods are the most accurate when the content of the image fulfills the ideal model they rely on. However, they lose accuracy in the presence of noise or when the linear structures are irregular and deviate from the ideal case. Machine learning techniques can alleviate this problem. However, they are mainly based on a classification framework. In this thesis, we show that classification is not the best formalism to solve the centerline detection problem. In fact, since the appearance of a centerline point is very similar to the points immediately next to it, the output of a classifier trained to detect centerlines presents low localization accuracy and double responses on the body of the linear structure. To solve this problem, we propose a regression-based formulation for centerline detection. We rely on the distance transform of the centerlines to automatically learn a function whose local maxima correspond to centerline points. The output of our method can be used to directly estimate the location of the centerline, by a simple Non-Maximum Suppression operation, or it can be used as input to a tracing pipeline to reconstruct the graph of the linear structure. In both cases, our method gives more accurate results than state-of-the-art techniques on challenging 2D and 3D datasets. Our method relies on features extracted by means of convolutional filters. In order to process large amount of data efficiently, we introduce a general filter bank approximation scheme. In particular, we show that a generic filter bank can be approximated by a linear combination of a smaller set of separable filters. Thanks to this method, we can greatly reduce the computation time of the convolutions, without loss of accuracy. Our approach is general, and we demonstrate its effectiveness by applying it to different Computer Vision problems, such as linear structure detection and image classification with Convolutional Neural Networks. We further improve our regression-based method for centerline detection by taking advantage of contextual image information. We adopt a multiscale iterative regression approach to efficiently include a large image context in our algorithm. Compared to previous approaches, we use context both in the spatial domain and in the radial one. In this way, our method is also able to return an accurate estimation of the radii of the linear structures. The idea of using regression can also be beneficial for solving other related Computer Vision problems. For example, we show an improvement compared to previous works when applying it to boundary and membrane detection. Finally, we focus on the particular geometric properties of the linear structures. We observe that most methods for detecting them treat each pixel independently and do not model the strong relation that exists between neighboring pixels. As a consequence, their output is geometrically inconsistent. In this thesis, we address this problem by considering the projection of the score map returned by our regressor onto the set of all geometrically admissible ground truth images. We propose an efficient patch-wise approximation scheme to compute the projection. Moreover, we provide conditions under which the projection is exact. We demonstrate the advantage of our method by applying it to four different problems.